Toybrick

标题: 多输入ONNX转换RK3588问题 [打印本页]

作者: HurryUp 时间: 2024-4-10 14:23
标题: 多输入ONNX转换RK3588问题
目前训练了一个多模态的分类模型，有两个非图片的多维度输入，分别是float32[1,32,256]和float32[1,3,16,114,114]。为了能够进行转化，将这两个维度重新导出为float32[1,1,32,256]和float32[1,48,114,114]的4维数据。设置为rknn.config(mean_values=[[123.675]], std_values=[[58.395]], target_platform=platform, optimization_level=0)；ret = rknn.load_onnx(model=MODEL_PATH, inputs=['physical','video'], input_size_list=[[1,1,32,256], [1,48,114,114]], outputs=['result'])；想请问针对这个模型这样设置是正确的吗？另外使用此设置在进行模型转化时报错

I expand_to_4d_conv: remove node = [], add node = ['Conv_41_0_expand0', 'Conv_41_0_expand1']E build: Catch exception when building RKNN model!E build: Traceback (most recent call last):E build: File "rknn/api/rknn_base.py", line 1546, in rknn.api.rknn_base.RKNNBase.buildE build: File "rknn/api/graph_optimizer.py", line 1276, in rknn.api.graph_optimizer.GraphOptimizer.fuse_opsE build: File "rknn/api/fuse_rules.py", line 5747, in rknn.api.fuse_rules._p_convert_global_avgpool_to_conv2E build: ValueError: need more than 3 values to unpackbuild model failed
请问该怎么解决

欢迎光临 Toybrick (https://t.rock-chips.com/)